Wavelet neural network for classification of transient signals

A method is presented for adaptively generating wavelet templates for pattern representation. The idea is to form “super-wavelets” that allows the shape of the wavelet to adapt to each presented pattern. The super-wavelets form compact features of the signal which can be dilated or translated to provide scale and position invariant classification. These wavelets will form a bank of filters that can be correlated with the input patterns. The correlation peaks would identify the patterns. We demonstrate the extraction of super-wavelets for speech signals and describe how this feature extraction technique can be employed as part of a classifier